Computational roadmap
System context
S2 is built around a custom Neuralink ASIC and has 1,680 electrode threads. It is designed mainly for stimulation, although it can also record. The threads can reach roughly 50 mm, which could provide access to the calcarine fissure and other deep parts of early visual cortex.
For each participant, we would need clinical notes, high-resolution anatomical and functional MRI, and the final thread locations. These data would support surgical planning and provide the starting point for an individual stimulation model.
The glasses would need a forward-facing camera and eye tracking. Eye position matters because the same object at a different retinal location may require a different stimulation pattern. Head pose may also need to be included.
Breakthrough Device designation does not itself allow human implantation. A first human study would still need an approved protocol and FDA authorization through an Investigational Device Exemption.
Core technical problem
Stimulate a small subset of sites and ask the participant what they perceive: location, size, brightness, shape, and stability. One electrode is not one pixel. Percepts may overlap, appear as broad flashes, and change when sites are stimulated together.
stimulation pattern → perceived visual outputThe forward model is participant-specific and must adapt to drift across days.
For a target such as an arrow, choose the channels, amplitude, timing, and duration that are most likely to produce it. The search has to account for interactions between sites and stay within fixed safety limits.
desired percept → stimulation across 1,680 channelsThe first deliverable is a small set of phosphene patterns that the participant can recognize reliably. Camera input can begin as a low-resolution binary map containing only the information needed for orientation and navigation.
Later versions can add depth and scene simplification while using the same participant-specific calibration loop.
Model optimization
Microsoft Research
I built a system in which AI agents write computational models of the visual cortex, score them against recorded neural activity, and use the result to decide what to try next.
The search is not restricted to a neural network or a predefined architecture. In a few hours, it tested thousands of models and found compact models that predicted early visual cortex better than the baselines we evaluated.
Blindsight
For each participant, the agents would search over forward models, stimulation rules, and the next calibration experiments to run. Each proposal would be scored using the participant’s actual reports and held-out trials.
As new calibration data comes in, the system keeps updating the participant’s model and testing better stimulation rules instead of choosing one model architecture at the start.